Facial Features Detection Using Adaptive Edge Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face detection algorithms for mobile devices require either dedicated hardware or strong CPU capabilities, making them unsuitable for low-cost mobile platforms without such resources, and struggle with uneven lighting, weak mouth edges, and false detections in textured regions.
Innovation Solution
A method that divides the skin tone area into parts, constructs luminance and edge maps, sets edge magnitude and absolute rejection thresholds, and extracts facial features using horizontal edges, reducing data processing and avoiding noise-related false detections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If face detection algorithms use dedicated hardware or strong CPU capabilities, then detection accuracy and real-time performance are improved, but device cost and complexity increase
Solution Approach 1:
The skin tone area is divided into a number of parts, and edge magnitude thresholds are calculated separately for each part. This segmentation allows the algorithm to handle uneven lighting conditions more effectively while maintaining computational efficiency on low-cost mobile platforms without dedicated hardware.
Solution Approach 2:
Different edge magnitude thresholds are applied to different parts of the skin tone area based on local lighting conditions. This local adaptation enables accurate edge detection in both well-lit and poorly-lit regions while maintaining overall detection accuracy on resource-constrained devices.
2Device complexity
If a single edge magnitude threshold is applied to the entire skin tone area, then processing is simplified, but detection accuracy deteriorates due to uneven lighting conditions
Solution Approach 1:
The skin tone area is divided into multiple parts, with each part having its own edge magnitude threshold calculated based on local statistics. This segmentation resolves the contradiction by maintaining relatively simple processing within each part while achieving high overall accuracy through localized threshold adaptation.
Solution Approach 2:
The edge magnitude threshold is made dynamic and adaptive to local lighting conditions rather than being fixed globally. Each part's threshold is calculated based on the statistical properties of edges in that specific region, allowing the system to adapt to uneven lighting without significantly increasing processing complexity.
3Measurement precision
If all edges above a low threshold are kept, then more facial features are captured, but false detections increase due to noise in untextured areas
Solution Approach 1:
Different edge magnitude thresholds are applied to different parts of the skin tone area based on local characteristics. This local quality approach allows the system to be sensitive to weak facial features in some regions while being more selective in regions prone to noise, thereby reducing false detections while maintaining feature completeness.
Solution Approach 2:
The edge magnitude threshold parameter is adjusted locally for different parts of the image based on statistical analysis of edge magnitudes in each region. This parameter adaptation enables the system to distinguish between meaningful facial features and noise by comparing edges against locally-appropriate thresholds rather than a global fixed threshold.
4Device complexity
If the mouth area is detected using the same threshold as eyes, then the algorithm is simpler, but mouth detection fails because mouth edges are generally weaker
Solution Approach 1:
The skin tone area is divided into parts that include separate regions for eyes and mouth. This segmentation allows the algorithm to calculate and apply different edge magnitude thresholds for eye regions versus mouth regions, accommodating the fact that mouth edges are generally weaker than eye edges while maintaining a relatively simple overall algorithm structure.
Solution Approach 2:
Different threshold criteria are applied to different facial feature regions based on their local characteristics. Eye regions use one threshold criterion while mouth regions use a different, more lenient criterion appropriate for weaker edges. This local quality approach enables accurate detection of both strong eye edges and weaker mouth edges without significantly increasing algorithm complexity.
Data Source
AI summary
There is described a method for facial features detection in a picture frame containing a skin tone area, comprising dividing (12) the skin tone area into a number of parts; and for each part of the skin tone area, constructing (14) a luminance map, constructing an edge map by extracting (18) edges from the luminance map, defining (20) an edge magnitude threshold, building (22) a binary map from the edge map by keeping only the edges having a magnitude beyond the defined edge magnitude threshold and eliminating the others; and then extracting (24) facial features from the built binary map. An inter-related facial features detector is further described.